The LAST R600 Rumours & Speculation Thread

Status
Not open for further replies.
Ailuros said:
How can you know what kind of room it leaves on R600 anyway? The 1xAA performance (yes it's 1x and not 0xAA dagnabit) could be by say 10% lower and the 4xAA score by 10% higher; that way you might have a significantly smaller drop from 1x to 4xAA, but the final result not being all that impressive either.
Click to expand...

Well I really don't see how an assumption that R600 is slower is any more valid than one that it's faster :???: For all we know it could be faster in 1xAA and take a lower 4xAA perf hit!

Would there be a high end TBDR in the works with a R600 alike arithmetic throughput yet exactly half it's bandwidth, would you say that it would lose against the latter because it has way less bandwidth?
Click to expand...

Of course not, but G80 isn't a TBDR, right? :) Efficiency matters yes but come on, do you really think efficiency differences between two IMR's is going to play a significant part with a 40+GB/s bandwidth discrepency? I really don't see how the assumption that R600 is less efficient holds any water either. What we expect is for R600 to have ridiculous amounts of bandwidth. All the assumptions that it's less efficient or can't make effective use of its extra bandwidth are just mitigating factors against the stomping of G80 that some expect!
 
trinibwoy said:
Well I really don't see how an assumption that R600 is slower is any more valid than one that it's faster :???: For all we know it could be faster in 1xAA and take a lower 4xAA perf hit!
Click to expand...

Pick your poison and lay your hand into an open fire for your choice then :p


Of course not, but G80 isn't a TBDR, right? :) Efficiency matters yes but come on, do you really think efficiency differences between two IMR's is going to play a significant part with a 40+GB/s bandwidth discrepency? I really don't see how the assumption that R600 is less efficient holds any water either. What we expect is for R600 to have ridiculous amounts of bandwidth. All the assumptions that it's less efficient or can't make effective use of its extra bandwidth are just mitigating factors against the stomping of G80 that some expect!
Click to expand...

I didn't say or imply that R600 is or might be less efficient. Amongst many other factors you need a specific amount of ROPs, TMUs and what not to make use of the exsessive bandwidth. All indications point so far at 16/8z ROPs for R600. No idea about TMUs to be honest, but if you attempt some reverse speculative math you have 24/8z ROPs@575MHz and 32 TMUs@575MHz on G80, which should show in today's scenarios severe bandwidth limitations.

My lucky guess is rather that ATI didn't want to take two steps like NV in order to go from 256 to 512bits.

Last but not least if I'd assume ~50% bandwidth it rarely translates into something even close to a 50% average increase if all other factors aren't increased by the same persentage. My only other point for the past few posts is that bandwidth alone won't be able to perform any wonders. Or if a living example would help, how much performance increase did the 1GHz ram bring the R580+ compared to the R580 and in which corner cases exactly?
 
Chalnoth said:
is infinitely more noticeable than blurring.
Click to expand...
I think that's a little bit simplistic. If you have a high level of anisotropic filtering enabled then sure, mip-maps are preferable to sparkle. But if you haven't got high-level AF then I, for one, find the presence of no-mipmap-sparkle much less objectionable than the combination of too-blurry textures and the "bow wave" effect that you see as the mipmap boundaries move along the ground ahead of you.
 
Ailuros said:
I didn't say or imply that R600 is or might be less efficient. Amongst many other factors you need a specific amount of ROPs, TMUs and what not to make use of the exsessive bandwidth. All indications point so far at 16/8z ROPs for R600. No idea about TMUs to be honest, but if you attempt some reverse speculative math you have 24/8z ROPs@575MHz and 32 TMUs@575MHz on G80, which should show in today's scenarios severe bandwidth limitations.
Click to expand...

Ah but we don't know what each ROP can do so sterile counts don't tell us much :p

My lucky guess is rather that ATI didn't want to take two steps like NV in order to go from 256 to 512bits.
Click to expand...

Could be but if rumours of 2Ghz+ GDDR4 are trrue then are they just going for the marketing bang?

Last but not least if I'd assume ~50% bandwidth it rarely translates into something even close to a 50% average increase if all other factors aren't increased by the same persentage. My only other point for the past few posts is that bandwidth alone won't be able to perform any wonders. Or if a living example would help, how much performance increase did the 1GHz ram bring the R580+ compared to the R580 and in which corner cases exactly?
Click to expand...

Yeah you're right, an R600 that isn't significantly faster than G80 in 1xAA probably won't benefit to the tune of its 50% bandwidth advantage once AA is turned on. R580+ didn't do much with its extra bandwidth IIRC. Hopefully R600 does something with all that bandwidth though.
 
trinibwoy said:
Ah but we don't know what each ROP can do so sterile counts don't tell us much :p
Click to expand...

Former newsblurb mentioned 2.5x times more Z-fillrate compared to R580.

Could be but if rumours of 2Ghz+ GDDR4 are trrue then are they just going for the marketing bang?
Click to expand...

You mean =/>1GHz (effective =/>2000 DDR)? :p Jokes aside why the heck not? I wouldn't be surprised to see utterly laughable creative math marketing stunts along that. And yes it's a perfectly logical answer to the creative math NV pushed with G80's specs.

Yeah you're right, an R600 that isn't significantly faster than G80 in 1xAA probably won't benefit to the tune of its 50% bandwidth advantage once AA is turned on. R580+ didn't do much with its extra bandwidth IIRC. Hopefully R600 does something with all that bandwidth though.
Click to expand...

Since G80 supports only up to 8xMSAA (the only real comparable mode should R600 support higher sample densities), that combined with 8x transparency supersampling should give the R600 always IMHLO a healthy boost ahead. All it takes is a healthy amount of alpha test textures in a test scene. And I'm only speculating in that direction because it's my impression that Supersampling eats more bandwidth in it's current form than anything else and isn't being compressed as MSAA.
 
Geo said:
Sometimes drivers don't just improve performance. Rys is gone for the day, but I suspect if he checks in tomorrow he's going to say that there was an issue with Cat 6.10 and AA in their built-in benchmark, and that in retrospect he should have mentioned it on the page. But you'll have to wait for him to do so.
Click to expand...
Yes, there was an issue with AA and the driver, so no AA was applied, and there's no in-game AF level control. I'm fairly sure I did mention that there was a bug, but it seems I didn't. Just an oversight, nothing suspicious.
 
There was an issue with AA and Company of Heroes on all graphics cards, and from what I can tell so far it still doesn't appear that AA is being applied (performance was impacted, though) with the latest drivers and game patch, at least not in the built-in Performance Test. I'll get a closer look at it when I can, and check out Xbit Labs' benchmark method.
 
Ailuros said:
With the only other difference that your quoted newsblurb is wrong on the ROP amount obviously.

R580:
16 * 4 z * 650MHz = 41600

41600 * 2.5 = 104000

16 * 8 z * 814MHz = 104192

The former thingy translated: http://translate.google.com/transla...&hl=de&ie=UTF-8&oe=UTF-8&prev=/language_tools

speaks of 5 operations per ALU/clock (4+1); sounds rather like Vec4+scalar to me.
Click to expand...

What am i missing here? R580 with quad-pumped ROPs?

What do you think on this: http://www.pcgameshardware.de/?menu=forum&s=thread&bid=421&tid=5451996&mid=5452060#5452060
A bit on the conservative side, but then - why would Ati waste so many transistors on their X1K test vehicle?
 
Jetboy said:
TheINQ says Quad GBU CrossFire.

http://www.theinquirer.net/default.aspx?article=37484
Click to expand...

You quoted the Inq in your first post? Damn...that's daring. :p

You meant quad GPU, right? Did anyone ever doubt this would happen? When all those lovely new motherboards started appearing the first thing I suspected was more videocards, or an additional card for the GPU/Physics processing (as we've all heard). Hope a reliable source can actually confirm QCF (Quad CrossFire?), and/or provide some more details.
 
Shader-ALUs
R580: 4 Quads á 12 ALUs (48)
R600: 8 Quads á 12 ALUs (96)
- deckt sich mit dem Äußerungen Ortons und den darauf basierenden Spekulationen.

Bandbreite
R580: 8x32 Bit (256)
R600: 16x32 Bit (512)
512 Bit sind ja mittlerweile beinahe gesetzt.

FP16-Filterleistung ("+700%")
R580 (650 MHz, 16 "Pipes") : 2.600 MTex/s (Shader)
R600 (650 MHz, 32 "Pipes") : 20.800 MTex/s (TMU)
Hier macht sich natürlich bezahlt, wenn man Funktionen nicht mehr emulieren muss, sondern die dafür gedachte Hardware nutzt.

Z-Fill (+150%)
R580 (650 MHz, 16 ROPs): 10.400 MZix/s
R600 (650 MHz, 20 ROPs): 26.000 MZix/s (double-z)
Click to expand...

Sounds legit to me? :cool:
 
Sound_Card said:
Sounds legit to me? :cool:
Click to expand...

Nope, i doesn't.

They keep referring to R600 as a simple doubling of the R580, but the R580 is not unified.
In addition to the 48 Pixel Shader ALU's, there are also 8 Vertex Shaders.

In a unified architecture, those "96" ALU's would have to be doing that additional work, not to mention Geometry Shading, new for the DX10 API.
 
Rebel44 said:
Hi I have UPDATE: It seems that tests from level505 were just confirmed link: http://www.ddworld.cz/graficke-karty/atir600-alias-ati-radeon-x2800xtx-testy-potvrzeny.html
Its in czech. BTW I know this guy and his info is most of the time very reliable - according to this info tests were done by someone from Microsoft with WEAKER:oops: :oops: version of X2800XTX on very early drivers which were modified by tester.
Click to expand...

Well, welcome to Beyond3D, assuming you aren't the author of that review out spamming. :smile:

So he's suggesting both that nvidia has a 8950GX2 product ready to go, and that AMD will be also introducing a GX2-like product called X2800XTX2?
 
Status
Not open for further replies.

Similar threads

A
The Official G84/G86 Rumours & Speculation Thread
22 23 24
Replies
460
Views
73K
Twinkie
T
Back
Top